gigaplan

name	gigaplan
description	Structured AI planning with cross-model critique. Use when the user says "/gigaplan", "make a plan", "structured plan", or wants rigorous planning with independent review.

Gigaplan Orchestration

You are orchestrating a gigaplan — a structured planning loop with cross-model critique.

Available Tools

Tool	Purpose
gigaplan_init	Initialize a plan directly when the agent needs to self-start gigaplan
gigaplan_doctor	Validate plan state, repair common JSON/output issues, and return recovery handoff
gigaplan_step	Get agent config for a step (task prompt, agent, output path)
gigaplan_advance	Process step output and advance the state machine
gigaplan_status	Check plan status
gigaplan_override	Manual intervention (add-note, abort, force-proceed, skip)

Use gigaplan_init for self-started flows. Do not rely on execute_command("/gigaplan ...") for agent-driven initialization. If gigaplan_advance fails on malformed JSON or a broken handoff, run gigaplan_doctor({ fix: true }) before retrying.

Orchestration Loop

For each step that needs an LLM:

1. Call gigaplan_step with planDir and step to get the agent config
2. Spawn the agent via agent_group with the returned config:

   agent_group({
     name: "Gigaplan: <step>",
     wait: true,
     agents: [{
       name: details.name,
       agent: details.agent,
       task: details.task,
     }]
   })
   

3. After the group completes, call gigaplan_advance with planDir and step
4. Read the result — it tells you the next step(s)

For evaluate and gate steps: skip the agent, just call gigaplan_advance directly.

Step Flow

clarify → plan → critique → evaluate
                                ↓
              CONTINUE → integrate → critique (loop)
              SKIP → gate → execute → review → done
              ESCALATE → ask user → override
              ABORT → done

Cross-Model Critique

The critique step SHOULD use a different model than the planning steps. This prevents self-bias — the planner never reviews its own work. The gigaplan_step tool handles agent routing automatically.

Key Rules

• Always advance after each agent completes — don't skip gigaplan_advance
• Evaluate is pure logic — no agent needed, just call gigaplan_advance
• Gate is pure logic — same as evaluate
• ESCALATE means ask the user — show them the evaluation and ask for an override decision
• Show progress — after each step, summarize what happened and what's next
• The plan lives in .gigaplan/ — all artifacts are there for auditability